Recurrence due to periodic multi-soliton fission in the defocusing nonlinear Schrodinger equation

Guo Deng1 , Sitai Li2 , Gino Biondini1,2 , and Stefano Trillo3

1 Department of Physics, State University of New York at Buffalo, Buffalo, New York 14260, USA
2 Department of Mathematics, State University of New York at Buffalo, Buffalo, New York 14260, USA
3 Department of Engineering, University of Ferrara, Via Saragat 1, 44122 Ferrara, Italy
(Dated: November 2, 2017)
We address the degree of universality of the Fermi-Pasta-Ulam recurrence induced by multisoliton fission
from a harmonic excitation by analysing the case of the semiclassical defocusing nonlinear Schrodinger
equation, which models nonlinear wave propagation in a variety of physical settings. Using a suitable
Wentzel-Kramers-Brillouin approach to the solution of the associated scattering problem we accurately predict,
in full analytical way, the number and the features (amplitude and velocity) of soliton-like excitations emerging
post-breaking, as a function of the dispersion smallness parameter. This also permits to predict and analyse
the near-recurrences, thereby inferring the universal character of the mechanism originally discovered for the
Korteweg-deVries equation. We show, however, that important differences exist between the two models,
arXiv:1711.00134v1 [nlin.SI] 31 Oct 2017

arising from the different scaling rules obeyed by the soliton velocities.

PACS numbers: 02.30.Ik, 05.45.Yv, 42.65.Sf, 47.20.-k

I. INTRODUCTION necessarily a prerogative of integrable models [11], the latter

The discovery by Fermi-Pasta-Ulam (FPU) that a low fre-
quency excitation of nonlinear oscillator chains give rise to re-
currence instead of equipartition, a phenomenon now known
as FPU recurrence [1, 2], turned into one of most consequen-
tial findings of nonlinear physics [37], leading to multiple
research avenues which are still actively investigated today
[811]. The first resolution of the apparent paradox was given
ten years later by Zabusky and Kruskal (ZK) [12] who showed
that in the continuum limit that gives rise to the weakly dis-
persive Korteweg-de Vries (KdV) equation [13], the recur-
rence can be understood in terms of solitons, that fission from
points of breaking (shocks) occurring in the periodic input
mode (a regime which is substantially confirmed, e.g. for
the quadratic, or so-called type, FPU chain [14]). Nearly
synchronous arrival of the solitons after interaction leads to
the recursion. An explicit estimate for the KdV recurrence
time was then given by Toda [15, 16]. Experimentally, the
recurrence from periodic input was reported in different sys-
tems ranging from electrical lattice networks [17] to continu-
ous wave systems such as ion acoustic plasma waves [18], and
gravity waves in shallow water [19], while it could be poten-
tially observed also for electron beams that exhibit KdV type
of breaking [20].
This result has led to the notion that a time scale exists for
which the dynamics of FPU is essentially integrable, that is, it
remains close to integrable limits, either the KdV in the con-
tinuum approximation or the discrete Toda chain [8, 21]. Over
much longer times the FPU chains eventually thermalise, as it
is now clear after decades of research on this topics [8, 9]. A
crucially important remaining issue, however, is to investigate
the metastable state characterised by the recurrences at inter-
mediate time scales. In this context, a natural question that has
remained surprisingly unaddressed is the degree of universal-
ity of the mechanism discovered by ZK for other integrable
models. While it is now accepted that FPU recurrence is not
constitutes a wide and extremely important class employed
to describe innumerous physical situations. Clearly, the im-
portance of integrable systems goes well beyond the specific
interest for FPU chains, since they often provide an accurate
description of several genuinely continuous physical systems.
In this context, breaking phenomena are a universal feature of
the weakly dispersive regime [22] (see also [23] for a recent
review) occurring for the nonlinear Schrodinger (NLS) equa-
tion [2433], naturally arising as the continuum limit of cubic
-FPU [4], the Benjamin-Ono equation [34, 35] and the Toda
lattice [3638], to name only a few. A special role in this
context is played by the defocusing or repulsive NLS equa-
tion, since it describes wave breaking phenomena recently ob-
served and analysed in areas as different as nonlinear optics
[2632], Bose-Einstein condensates [24, 25], and spin waves
[33]. Furthermore, the periodic input problem for such model
was recently demonstrated to be accessible in fiber optics ex-
periments, which show fission of dark solitons from peri-
odic points of breaking [29, 30]. Yet, natural and fundamen-
tal questions concerning how to predict the number and the
features of the solitons and whether recurrence should be ex-
pected remain to date completely open in such model.
The purpose of this paper is to give an answer to these ques-
tions and to demonstrate that soliton generation as a result of
hydrodynamic-like instabilities, and the resulting FPU-like re-
currences, are in fact a more general feature of nonlinear wave
evolution equations with small dispersion. To this end, we
show that such phenomena can be effectively treated analyti-
cally using the scattering problem and its finite gap formula-
tion associated with the defocusing NLS equation. While our
results significantly broaden the universality of the ZK mecha-
nism of recurrence, they allow to clarify that the phenomenon
presents, for the NLS equation, considerable differences with
respect to the case of the KdV equation.
We remark that FPU recurrences have also been referred to
in the literature in the focusing regime of the NLS equation.

These, however, are exact (instead of near) recurrences that
occur through a different kind of mechanism involving mod-
ulation instability of a strong background (a forbidden regime
for the defocusing NLS equation which is well known to be
modulationally stable). Such scenarios have been investigated
in fluids and optics only in a regime that involve few dominant
Fourier modes, i.e. far from the weakly dispersive limit that
we are interested in [3944]. Conversely, wavepackets with-
out background in the weakly dispersing limit of the focusing
regime are known to exhibit a different type of breaking (i.e.,
elliptic unbilic catastrophe [45]) compared with the defocus-
ing case. The complicated dynamics in the evolution stage
beyond the catastrophe cannot be reduced, in general, to the
fission of solitons that simply move apart with different ve-
locities [46, 47]. Moreover the global scenario for a periodic
input mode in the focusing regime was never addressed so far,
and will require a specific analysis that will be proposed else-
where.
Motivated by the above discussion, here we consider the
following dimensionless periodic initial value problem for the
defocusing NLS equation
ieqt + e2 q xx 2|q|2 q = 0 , (1)
with initial conditions
q( x, 0) = cos x , 6 x 6 . (2)
The variables x and t are typically space and time, respec-
tively (e.g., see [48, 49]), though in nonlinear fiber optics they
have the reversed role of retarded time and propagation dis-
tance [29, 30, 50], and e quantifies the smallness of dispersion
[29] (in quantum-mechanical settings e is also proportional to
Plancks constant h [24]) compared to the strength of nonlin-
ear effects. Without loss of generality (taking advantage of the
scaling invariances of Eq. (1)), we have normalized the period
X p of the cosine initial value to 2.
The nonlinearity in Eq. (1) induces conversion towards
high-frequency modes, i.e. odd harmonics m/X p , m =
3, 5, . . . of the m = 1 input frequencies, a phenomenon com-
monly known in nonlinear optics as multiple four-wave mix-
ing (mFWM) [30, 51, 52]. In the regime of interest here,
e  1, mFWM becomes very efficient and causes strong
steepening of the cosine fronts, until breaking (hydrodynamic
instability [29]) occurs followed by fission into soliton-like
excitations, as shown by the numerical simulation in Fig.
1. Marked differences with the corresponding phenomena in
the KdV equation [12, 19] are the degenerate mechanism for
breaking that occurs at the null (vacuum) points (cf. [28, 31],
where similar mechanism is analysed in detail for dark type
input on a constant background) and the fission into pairs with
opposite velocities, which reflect the bidirectional (NLS) ver-
sus unidirectional (KdV) dispersive hydrodynamic nature of
such models [23]. Nevertheless, even in the NLS equation,
solitons eventually give rise to a near-recurrence as shown in
Fig. 1. Our goal, here, is to provide a quantitative descrip-
tion of this phenomenon, including analytical estimates of the
2
number of fissioning solitons and their velocities as a function
of e, and then use these results to characterise the recurrent
behavior.
The outline of this work is the following. In section II
we characterize the spectrum of the NLS equation (1) in the
small dispersion limit with initial conditions (2) using a suit-
able WKB expansion. In section III we study the properties
of the effective solitons arising from the corresponding spec-
trum. In section IV we discuss the recurrence of initial condi-
tions. Section V offers a few concluding remarks. The details
of the calculations are given in the Appendix.
II. NLS SPECTRUM IN THE SMALL DISPERSION LIMIT
II.1 Scattering problem and monodromy matrix
Since the NLS Eq. (1) is completely integrable, the initial
value problem can be solved via the inverse scattering trans-
form (IST). In particular, the relevant formalism here is the
IST with periodic boundary conditions, or finite-gap theory
[5356]. According to the periodic IST, the nonlinear ex-
citations embedded in the initial datum are encoded in the
spectrum of the scattering problem associated with the NLS
equation (1), i.e. the well-known Zakharov-Shabat (ZS) with
cosine potential
ex = (ik3 + cos x 1 ) , (3)
where ( x, k ) = (1 , 2 ) T is the vector eigenfunction, and
1 and 3 are the first and third Pauli matrices, respectively.
Since the scattering problem (3) is self-adjoint, all eigenvalues
k are real. Applying the change of variables v = 1 + 2
and v = 1 2 , Eq. (3) can be reduced to the second-order
ordinary differential equation,
e2 v xx + (cos2 x e sin x ) v = v , (4)
which is the time-independent Schrodinger equation with an
e-dependent potential and eigenvalue = k2 > 0. Since
the potential is periodic, Bloch-Floquet theory can be used to
show that Eq. (4) admits bounded solutions if and only if
2 6 trM 6 2 , (5)
where M () is the monodromy matrix of the problem, de-
fined as M = Y 1 ( )Y ( ), and Y ( x ) is any fundamental
matrix solution of Eq. (4). The values of for which Eq. (5) is
satisfied comprise the spectrum of Eq. (4). Note that, toeach
nonzero value of , there correspond two values k = in
the original scattering problem (3).
II.2 Wentzel-Kramers-Brillouin (WKB) analysis
Since no solutions in closed form are available for Eq. (4),
we apply the WKB method (e.g., see [57]) to obtain asymp-
totic expansions for the solutions of Eq. (4) and therefore for
 3

0
x

0.5

-2
0
0 2 4 6 8 10 12 14 16
t

FIG. 1. Density plot of |q( x, t)| from numerical solution of the initial value problem in Eqs. (12) with e = 0.1, showing the fission into
soliton-like excitations as a result of breaking (hydrodynamic-like instability, see early stage t < 1) occurring in the null points of the input
cosine and the following near-recurrence at t ' 15.5.

trM in the dispersionless limit, similarly to the approach de-
veloped in Refs. [19, 58]. It is convenient to introduce the
shorthand notation
Q( x ) = cos2 x .
For > 1, no turning points are present, and straightforward
calculations (see Appendix A.1 for details) yield

variables v and v, and can be solved exactly to get 12 trM = 1,
trM = 2 cos S0 ()/e , > 1, (7a)
R p
where S0 () = Q( x ) dx. On the other hand, for
0 < < 1 four turning points are present, and the WKB
analysis is considerably more complicated. In this case, the
main difficulty comes from the fact that one must first con-
struct asymptotic solutions in each of the regions away from
the turning points and near each turning points, and then
match the solutions in the transition regions around the turn-
ing points, obtaining so-called connection formulae which al-
low one to continue the asymptotic expressions for the eigen-
functions over the whole spatial domain. These expressions
can then be used to finally construct the monodromy matrix.
Omitting the details for brevity (see Appendix A.2 for de-
tails), we find the following expression for the trace of the
monodromy matrix
< 1 is divided into alternating bands and gaps correspond-
ing to the values of for which 1 6 12 trM 6 1 (bands)
or 12 trM < 1 (gaps), respectively. The bands and gaps are
(6) separated by a sequence of band edges n for n = 0, 1, 2 . . . ,
which are the values of such that 12 trM = 1. Note that the
value = 0 is always part of the spectrum (this can be seen by
noting that for k = 0 the scattering problem decouples in the
which coincides with the limiting value of Eq. (7b) as k 0).
As an example, in Fig. 2 (left panel) we show the depen-
dence of 12 trM on the spectral parameter for e = 0.15.
Note that, since trM exhibits exponentially large oscillations,
to capture the whole behavior in a single plot we plot the
quantity f (trM/2) instead of 12 trM itself, as in [59], with
f () defined as f () = for || 6 1 and f () =
sign() (1 + log10 ||) for || > 1. In such figure we
compare the above WKB asymptotic expressions for 12 trM in
Eq. (7) with the values obtained from direct numerical integra-
tion of Eq. (4). As can be seen from the figure, the difference
between the asymptotic expressions and the numerical values
is negligible for our purposes.
III. EFFECTIVE SOLITONS

trM = III.1 Asymptotic properties of the spectrum

2 4 sin2 (S1 ()/e) cosh2 [S2,e ()/e], 0 < < 1,
(7b) Equation (7b) implies that, for 0 6 < 1, each band is
clustered around a maximum of 12 trM, at which 12 trM () =
where S2,e () = S2 () + e log 2, 1, and that the n-th maximum = zn is given by the solution
Z c q of
S1 ( ) = Q( x ) dx , (8a)
c
Z c q S1 (zn ) = ne . (9)
S2 ( ) = | Q( x )| dx , (8b)
c Since = k2 , all spectral bands of the scattering problem (4)
correspond to a symmetric pair of nonlinear excitations of
and c = arccos( ). the NLS Eq. (1) except for the first spectral band, centered
In the following, we conveniently refer to the half trace around k = 0, which generates just one nonlinear excitation.
1
2 trM, which, according to Eqs. (7), turns out to be bounded Hence, the number Ne of nonlinear excitations of the prob-
by the unit value, i.e. 12 trM 6 1, for all > 0. In particular, lem is related to the total number Nb of spectral bands in the
Eq. (7a) shows that the range 1 < < forms one infinitely range 0 6 < 1 by Ne = 2Nb 1. Since S1 (1) = 2,
long band. Conversely, Eq. (7b) shows that the range 0 6 Eq. (9) then yields Nb = b2/(e)c + 1 and therefore Ne =
 4

defined by
0 1
10
S2 (s ) = e log[2/( )]. (11)
f(trM/2)

-5

Nm
-10
Denoting by Nm the total number of maxima in [0, s ), we
100
have that Nm is also the number of distinct soliton pairs (in
-15
0 0.5 1 1.5 2 2.5 10-1 which any two symmetric gray solitons are counted as one,
like the lone black soliton). In light of the above discussion,
from Eq. (9) we then have
FIG. 2. Left panel: Half trace of the monodromy matrix 12 trM as a
function of the eigenvalue with e = 0.15, comparing the asymp-
Nm = bS1 (s )/(e) + 1c . (12)
totic expressions (7) (dashed red) with the results from numerical in-
tegration of the scattering problem (4) (solid blue). Right panel: the
number Nm of solitonic excitations as a function of e. Red curve: The total number of solitons present in the problem is then
full WKB prediction, Eq. (12). Gray curve: linear approximation, simply Ns = 2Nm 1. As shown in Fig. 2 (right panel),
Eq. (13a). Blue squares: the value obtained from numerical integra- the above WKB estimate for the total number of solitons is in
tion of Eq. (3). Orange stars: same from direct numerical integration excellent agreement with results from numerical integration
of Eq. (1). Also shown for comparison (magenta curve) is the total
of Eq. (3) and also in good agreement with direct numerical
number of excitations Ne .
simulations of Eq. (1) (again with the caveat regarding the
difficulty of counting numerically the number of excitations
in case of small dispersion).
2b2/(e)c + 1. It was conjectured in [29, 30] that these ex-
citations display soliton-like behavior. For a fixed value of e, We can also derive fully explicit approximations for the
however, only some of the excitations in the problem resem- number of solitons as a function of e by considering either
ble the dark solitons of the defocusing NLS equation. Note a linear or quadratic expansion of S1 () near = 0 and a
however that counting the number of excitations from direct linear expansion of S2 () near = 1 (see Appendix A.3 for
numerical simulations of the NLS equation presents two ma- details), thus obtaining
jor challenges. The first one is that not all excitations might be  
identifiable in the output. The second one is that it is highly 1 1 2
Nm,linear = log +1 , (13a)
nontrivial to distinguish which ones among all visible excita- 2e
tions are of solitonic or non-solitonic type. In order to dis-  
4 2+1 2 2+1 2
tinguish soliton-like excitations from non-solitonic ones, one Nm,quadratic = log +1 ,
8 2e 2 2
must look at the the relative width Wn = wn /(wn + gn ) of
(13b)
the n-th band, where wn (gn ) is the width of n-th band (adja-
cent gap). By expanding the expression for trM in Eq. (7b),
we find (see Appendix A.3 for details) respectively. Corresponding approximations for Ns follow ac-
cordingly. The values of Nm,linear are shown in Fig. 2; those
2 of Nm,quadratic are almost indistinguishable from the exact
Wn = exp(S2 (zn )/e) + O(eeS2 (zn )/e ) . (10) count given by Eq. (12). Note that in both cases the leading

order term is independent of the specific value chosen for the
threshold .
The proper solitonic limit of each excitation is obtained
when the relative bandwidth Wn tends to zero. Of course,
since the bandwidths are always greater than zero, one never
has true solitons in the periodic problem. (This is at vari- 2 1
ance, for instance, with what happens when one considers 0.8
1.5
the dark potential q( x, t = 0) = tanh x on the infinite line,
0.6
which is reflectionless for integer 1/e, containing a number
|V |
n

1
A

Ns = 2/e 1 of discrete eigenvalues that correspond to true 0.4

solitons, as shown in [60].) Nonetheless, excitations asso- 0.5
0.2
ciated with very narrow relative bandwidths (Wn  1) be-
come closer and closer approximations of the dark solitons of 0 1 2 3 4 5 6 7 0 1 2 3 4
n n
the defocusing NLS equation. Correspondingly, given a fixed
threshold  1, similarly to [19, 58] we define a nonlinear FIG. 3. Absolute value of the soliton velocity |Vn | (left panel) and
excitation of the periodic problem to be an effective soliton if amplitude An (right panel) for the soliton set associated to a given
the band to which it is associated is such that Wn < . Us- index n. Red circles: asymptotic values; Blue dots: direct numerical
ing Eq. (10) and solving this inequality, we then see that the simulations of Eqs. (1)-(2). Here e = 0.06, resulting in Ns = 15 and
solitonic bands are confined to the range [0, s ), where s is q = 0.898.

1 1
0.8 0.8
|q(x,t)|
|q(x,t)|
0.6 0.6
0.4 0.4
0.2 0.2
-2 0 2 -2 0 2
x x
FIG. 4. Near-recurrence of initial condition for e = 0.3 (left panel)
and e = 0.08 (right panel), comparing the initial value (2) (solid
black) and q( x, t) at the near-recurrence time (dashed red).
III.2 Amplitudes and velocities of the effective solitons
Recall that the single dark-soliton solution of the defo-
cusing NLS equation associated with a discrete eigenvalue
k = k o is such that
|q( x, t)|2 = k2o
+ (q2 k2o ) tanh2 [
p
q2 k2o ( x xo 2k o t)] , (14)
where q = limx |q( x, t)| is the background ampli-
tude. Also recall that, to each solitonic band centered at
= zn 6= 0, with zn defined by Eq. (9), there corresponds a
pair of dark solitons with discrete eigenvalues k n = zn .
The band centered at z0 = 0, instead, corresponds to a sin-
gle black soliton. We then obtain the velocity and the am-
plitude of the n-th set of solitons as Vn = 2 zn and
An = q2 zn , respectively, n = 0, 1, 2, . . . Finally, recall
that, in the infinite line problem, the continuous spectrum is
k (, q ) (q , ). Identifying the beginning of
the continuous spectrum with the first non-solitonic band, we
therefore have q = z Nm +1 . Figure 3 shows a comparison
between the above WKB estimates for the soliton velocities
and amplitudes and the values obtained from direct numerical
simulations of the NLS equation.
IV. RECURRENCE OF INITIAL CONDITIONS
Similarly to the KdV equation, the evolution ruled by the
defocusing NLS equation is expected to nearly recover the
cosine initial value when the solitons simultaneously return
to their initial location after traveling an integer number of
periods.
Figure 4 shows the solution at the (numerically determined)
recurrence time for e = 0.3 and e = 0.08. The numerically
determined dependence of the recurrence time on e is also
shown in Fig. 5. Next we show how one can use the above
WKB results to obtain an estimate of the recurrence time,
where we neglect the small shifts due to soliton collisions.
Since our purpose is also to study the differences between
recurrences in the KdV and the NLS equations, let us briefly
5
consider first the case of the KdV, which we write as
ut + 6uu x + e2 u xxx = 0 . (15)
An estimate for the recurrence time in Eq. (15) can be ob-
tained from the WKB analysis developed in Refs. [19, 58].
Adopting the same notation of such references, we recall that,
in this case, the solitonic bands correspond to nonlinear exci-
tations that are close to bright KdV solitons of the form
un ( x, t) = u0 + An sech2
p 
An /2( x Vn t)/e , (16)
where An is the amplitude of the n-th soliton from the back-
ground level u0 , and Vn = 6u0 + 2An is the correspond-
ing velocity. The amplitude can be calculated as [58, 59]
An = 2(ref n ) for n = 1, 2, . . . N, where N is the
number of the solitonic bands and ref = u0 corresponds
to the ( N + 1)-th band, or first non-solitonic band. Accord-
ing to the analysis in Ref. [58] and adopting a linear approx-
imation for the eigenvalues, we find that the amplitudes and
hence the velocities scales linearly with the soliton order n,
with increment A = An An+1 = 2 2e and hence
V = Vn Vn+1 = 4 2e. Clearly, this means that after
a time TKdV such that V TKdV = 2, all the solitons recur
having traveled a multiple of the period (recall that 2 is the
period of the cosine, and note that, obviously, the contribution
to the velocity that is common to all solitons does not affect
the recurrence time). From this condition we then immedi-
ately find
TKdV = . (17)
2 2e
Remarkably, the estimate in Eq. (17) coincides with an earlier
one by Toda [15, 16] (once one performs the trivial rescalings
of the spatial and temporal variables so as to obtain the same
form of the KdV equation). We note, however, that Todas es-
timate was obtained using a quadratic (parabolic) approxima-
tion for the cosine potential, which amounts to approximating
the eigenvalues with those of an harmonic oscillator. Impor-
tantly, numerical results (not shown) based on the integration
of the KdV equation confirm the scaling of the recurrence time
with 1 from Eq. (17), while the recurrence times from the
numerics turn out to be slightly overestimated by the formula,
possibly due to the soliton-soliton interactions, which are ne-
glected in both our and Todas approaches.
Returning back to the defocusing NLS equation, the time
needed for the n-th soliton pair to travel a distance equal to a
whole period is Tn = 2/|Vn |, with Vn = 2 zn as be-
fore. By employing again a linear expansion of S1 () around
= 0 (see Eqs. (35a) in Appendix), we obtain the explicit
expression Vn = 2 2ne, which in turn yields
Tn = . (18)
2ne
This result shows that the key difference between the NLS and
KdV equations is that, for the latter, Vn is linearly proportional

to the soliton index n (in the WKB limit), whereas the above
expression shows that for the NLS solitons Vn depends on
the square root of n. This difference is reflected in the travel
time Tn . Therefore, while in the KdV equation all solitons
simultaneously return to their initial position at the recurrence
time, this is not true in the NLS equation, and the situation is
more complicated in this case.
More precisely, neglecting the interaction-induced position
shifts, for any fixed integer m > 0, all soliton pairs with
indices n = m, 4m, 9m . . . , l 2 m, . . . (with l any fixed posi-
tive integer) will return to their initial position, which is also
the position of the stationary black soliton, at integer multi-
ples of the recurrence time t Rm = / 2me. The value
of t R1 as a function of e, shown in Fig. 5, is in good agree-
ment with the numerically determined recurrence times. Im-
portantly, note that the recurrence time for the NLS equation
scales like e1/2 , whereas the recurrence time for the KdV
equation scales like e1
In order to further characterise the degree of the recurrence,
we introduce the following figure of merit
R = 1 k|q( x, t)| | cos x |k/k cos x k , (19)
2 in fiber optics (in the context of mFWM dynamics) and possi-
where k f k2 = 0 | f ( x )|2 dx quantifies the energy con-
R
tained in a periodic signal. (Note that k cos x k2 = .) The
value of R is a measure of the strength of the recurrence.
A perfect recurrence corresponds to R = 1. The numeri-
cally determined dependence of R on e is reported in Fig. 5,
showing a strong deterioration for decreasing e, whereas near-
perfect recurrence is achieved for relatively large values of e.
The latter fact is straightforwardly explained from Eq. (12),
which implies that, for e > 0.163, there is only one soliton
pair besides the stationary black soliton. In this regime, the
initial condition
is recovered almost exactly at multiples of
t R = /( 2e), as also illustrated in Fig. 4 (left panel), con-
sistently with earlier experiments [61, 62] (se also [29]). How-
ever, for e < 0.163, multiple soliton pairs come into play, and
the recurrence becomes progressively worse for decreasing e,
as illustrated for e = 0.08 in Fig. 4 (right panel). Importantly,
a similar calculation for the KdV equation also shows a re-
currence degradation with decreasing e. However, the scaling
law in the two models is different. Specifically, R scales pro-
portionally to e for the KdV, and as e1/4 for the NLS (see
Fig. 4, right panel). No analytical results are available in ei-
ther model which explain this dependence.
V. CONCLUSIONS
In summary, we have shown that, in systems governed by
the weakly dispersing defocusing NLS equation, the fission
of solitons from a periodic wave and their recurrence can be
described in full analytical fashion. The result demonstrates
that the phenomenon of near-recurrence discovered by ZK is
observable in other integrable systems, though important dif-
ferences arise due to the different scaling of soliton veloci-
ties with their order. We believe that our approach can be
6
100
recurrence strength
101
recurrence time
10-1 100 10-1 100
FIG. 5. Left panel: recurrence time as a function of e, as computed
from direct numerical simulations of Eqs. (1)-(2) (black circles),
along with a linear fit of the numerical data (red line) and the analyt-
ical prediction trecur = t R1 (dashed blue line). Right panel: Figure
of merit of recurrence strength [Eq. (19)], as computed from direct
numerical simulations (black circles), and its linear fit for e < 0.3
(red line).
easily extended to other integrable systems (e.g., such as the
Benjamin-Ono equation) in order to produce a critical assess-
ment on the general degree of universality of the recurrence
phenomenon originally discussed by ZK for the KdV equa-
tion. We also expect our results to be experimentally verifiable
bly in other settings (Bose-Einstein condensates, spin waves,
oscillator chains).
The results of this work also open up the interesting ques-
tion of whether a similar approach can also be used to study
the small dispersion limit of the focusing NLS equation with
periodic boundary conditions. Of course the focusing case
is expected to be more challenging than the defocusing one.
From a mathematical point of view, this is because the cor-
responding spectral problem is no longer self-adjoint, which
means that the spectral bands are not restriced to the real
axis, and which puts into question whether one can effectively
make use of the WKB method. Physically, the problem is also
expected to give rise to more complex phenomena because of
the presence of modulational instability, which becomes more
and more severe in the dispersionless limit (i.e., as e 0).
Acknowledgment. We thank A. Armaroli for insightful
discussions during the early stage of this work. This work
was supported in part by the National Science Foundation un-
der grant numbers DMS-1614623 and DMS-1615524. S.T.
acknowledges support from the Italian Ministry of University
and Research (PRIN-2012BFNWZ2)
APPENDIX
Here we provide the details on the analytical calculation of
the monodromy matrix and the relative bandwidths.
A.1 The WKB expansion and the range > 1
The WKB expansion. Recall that the scattering problem is
given by Eq. (4). We look for an asymptotic expansion for the
 7

solution v of Eq. (4) in the form (a) Region 1, x [, c2 );

(b) First transition region, comprised by a neighborhood of
v( x ) = [ A( x ) + O(e)] eiS( x)/e , (20)
x = c2 ;
as e 0. Substituting Eq. (20) into Eq. (4), one obtains, at (c) Region 2, x (c2 , c1 );
the first two orders in the expansion, the eikonal equation and (d) Second transition region, comprised by a neighborhood
the transport equation, i.e., of x = c1 ;
S2x = Q( x ), (21a) (e) Region 3, x (c1 , c1 );
(f) Third transition region, comprised by a neighborhood
2iSx A x + iSxx A + sin x A = 0, (21b)
of x = c1 ;
respectively. These equations are readily integrated to obtain (g) Region 4, x (c1 , c2 );
Rq (h) Fourth transition region, comprised by a neighborhood
S ( x ) = Q( x ) dx , (22a) of x = c2 ;
(i) Region 5, x (c2 , ].
rq q
4
A ( x ) = Q( x ) i cos x Q( x ) , (22b)
The various sub-regions are illustrated in Fig. 6. Next, we
discuss the WKB solution of the scattering problem in the var-
up to arbitrary additive and multiplicative constants, respec- ious sub-regions.
tively, where Q( x ) = cos2 x was defined in Eq. (6), and
A corresponds to the plus/minus sign in S ( x ), respectively. (a) In region 1, x [, c2 ), the WKB approximation
Observe that, for > 1, Q( x ) > 0 for all x R. Con- for the general solution of the scattering problem (4) is
versely, for 0 < < 1 the range x [, ] divides into v1 ( x ) = a1+ v1+ ( x ) + a1 v1 ( x ), (26)
three sub-domains:
where
(i) Q( x ) < 0 for x [, c2 ) (c1 , c1 ) (c2 , ],
S S
Rx q
(ii) Q( x ) > 0 for x (c2 , c1 ) (c1 , c2 ), | Q( x )|dx/e),
S
v1 ( x ) = A exp( (27)
c2
(iii) Q( x ) = 0 for x = c1 , c2 ,

where c1 = arccos( ) and c2 = c1 (with c = c1 ). We and
then need to study the WKB solutions in these two ranges of 1/2 
A = | Q( x )|1/2 q | Q( x )|1/4 .

separately. (28)
Trace of the monodromy matrix for > 1. For any value
of in this range, Q( x ) is positive. Two linearly independent (b) In the first transition region, x (c2 , c2 + )
solutions are given, in the WKB approximation, by with > 0, Eq. (4) becomes
p p
v ( x ) = A ( x ) eiS ( x)/e , (23) e2 v xx + (2k 1 k2 ( x + c2 ) e 1 k2 )v = 0,

an asymptotic expansion for v in the first transition region is

with A ( x ) and S ( x ) given by Eq. (22). We can write a
given by
corresponding fundamental matrix solution of the first-order
system associated with Eq. (4) as a Wronskian: v12 ( x ) = c1 Ai[ ( x )] + c1+ Bi[ ( x )],
Y ( x ) = Wr(v , v+ ) . (24) where
p
Since Q( x ) 6= 0 in this case, such a fundamental matrix solu- ( x ) = a1/3 ( x + c2 e/2k)/e2/3 , a = 2k 1 k2 .
tion is valid over the whole range x [, ]. Therefore the
mondoromy matrix can be obtained simply as (c) In region 2, x (c2 , c1 ), the WKB approximation
to the general solution has two equivalent expressions
M = Y 1 ( )Y ( ) . (25)
Rx q
v2 ( x ) = a2+ A+ ( x ) exp(i | Q( x )|dx/e)
Straightforward calculations then show that the trace of M is c2
given by Eq. (7a).

Region1 Region2 Region3 Region4 Region5

A.2 Trace of the monodromy matrix for 0 < < 1

-c2 -c1 c1 c2 x
WKB solutions for 0 < < 1. For all values of in -p p
this range, there are four turning points, namely c1 , c2 . First Second Third Fourth
Correspondingly we need to discuss the behavior of the eigen-
functions in the following nine sub-regions of the fundamental FIG. 6. The sub-regions of the fundamental domain x [, ] for
the WKB analysis in the range 0 < < 1.
period x [, ]:
 8

Rx q with A given by Eq. (22).

+ a2 A ( x ) exp(i | Q( x )|dx/e), (29a)
c2 (h) In the fourth transition region, x (c2 , c2 + ),
Rx q Eq. (4) becomes
v2 ( x ) = a2+ A+ ( x ) exp(i | Q( x )|dx/e) p p
c1 e2 v xx (2k 1 k2 ( x c2 ) e 1 k2 )v = 0,
Rx q
+ a2 A ( x ) exp(i | Q( x )|dx/e), (29b) an asymptotic expansion for v in the fourth transition region
c1 is given by
with A given by Eq. (22). v45 ( x ) = c4 Ai[( x )] + c4+ Bi[( x )],
(d) In the second transition region, i.e., x (c1
, c1 + ), Eq. (4) becomes where ( x ) = a1/3 ( x c2 e/2k)/e2/3 .
(i) In region 5, x (c2 , ], the general solution in the
p p
e2 v xx (2k 1 k2 ( x + c1 ) + e 1 k2 )v = 0 .
WKB approximation is
An asymptotic expansion for v in the second transition region Rx q
is given by v5 ( x ) = a5+ A+ ( x ) exp( | Q( x )|dx/e)
c2
v23 ( x ) = c2 Ai[ ( x )] + c2+ Bi[ ( x )],
Rx q
+ a5 A ( x ) exp( | Q( x )|dx/e),
where ( x ) = a1/3 ( x + c1 + e/2k)/e2/3 . c2
(e) In region 3, x (c1 , c1 ), the general solution in the
with A given by Eq. (28).
WKB approximation also has two equivalent expressions
Connection formulae for 0 < < 1. Having computed
Rx q WKB expressions for the solution of the scattering problem
v3 ( x ) = a3+ A+ ( x ) exp( | Q( x )|dx/e)
c1
in the various sub-domains for 0 < < 1, we now need
to match the solutions across different sub-regions. First we
Rx q
+ a3 A ( x ) exp( | Q( x )|dx/e), (30a) match v1 ( x ) with v12 ( x ). Explicitly,
c1
1 2 3/2
( a1+ e 3 a| x+c2 | /e
4
Rx q v1 ( x ) = p
v3 ( x ) = a3+ A+ ( x ) exp( | Q( x )|dx/e) 4
a | x + c2 |
c1 32

a| x +c2 |3/2 /e
+ a1
4
Rx q e )(1 + O(| x + c2 |)),
+ a3 A ( x ) exp( | Q( x )|dx/e), (30b) x c2 ,
c1
1 2 3/2 2 3/2
1/2  v 12 ( x ) = p
4
( 21 c1 e 3 + c1+ e 3 )
2
| Q( x )|1/2 | Q( x )|1/4 .

where A = +q
(f) In the third transition region, x (c1 , c1 + ), (1 + O(1/ 3/2 )), .
Eq. (4) becomes
p p Requiring that these two expansions match, we obtain the con-
e2 v xx + (2k 1 k2 ( x c1 ) + e 1 k2 )v = 0 . nection formula
    4
An asymptotic expansion for v in the third transition region is 2 2 0
 
c1 a1
= C1 , C 1 = .
given by c1+ a1+ ( ae)1/6 0 1
v34 ( x ) = c3 Ai[z( x )] + c3+ Bi[z( x )],
To match v2 ( x ) with v12 ( x ), note that
where z( x ) = a1/3 ( x c1 + e/2k)/e2/3 . 1 2 3/2
4
(g) In Region 4, x (c1 , c2 ), the WKB approximation to v2 ( x ) = p
4
( a2+ ei/4 e 3 i a(x+c2 ) /e
a ( x + c2 )
the general solution again has two equivalent expressions
i/4 23 i a( x +c2 )3/2 /e
+ a2
4
e e )(1 + O( x + c2 )),
Rx q
v4 ( x ) = a4+ A+ ( x ) exp(i | Q( x )|dx/e) x c2 ,
c1
1
c1 sin( 32 | |3/2 +

Rx q v 12 ( x ) = p 4)
+ a4 A ( x ) exp(i | Q( x )|dx/e), (31a) 4
2 | |
c1
+ c1+ cos( 23 | |3/2 + 4 ) (1 + O(1/| |3/2 )), ,

Rx q
v4 ( x ) = a4+ A+ ( x ) exp(i | Q( x )|dx/e)
c2 implying
Rx q  +
a2
 
c1 ( ae)1/6 i 1
 
+ a4 A ( x ) exp(i | Q( x )|dx/e), (31b) = C , C = .
c2 a2 2
c1+ 2
2 2 i 1
4
 9

Note also that a2 in Eq. (29b) relates to a2 in Eq. (29a) by 1 2 3/2 2 3/2
v 34 ( x ) =
4
( 12 c3 e 3 z + c3+ e 3 z )
 +  + 2
z
a2 a
= C3 2 , C3 = ei3 S1 ()/e , (1 + O(1/z3/2 )), z .
a2 a2
Requiring that these two expansions match, we obtain the con-
with
nection formula
Rc2 q    
S1 ( ) = Q( x )dx. (32) c3 a3
+ = C1 .
c1 c3 a3+
A plot of S1 () is shown in Fig. 7. Next we match v4 ( x ) with v34 ( x ), note that
Next we match v2 ( x ) with v23 ( x ). Note that 1 i/4 2 ia(xc )|3/2 /e
+ 4
i/4 2 ia| x+c |3/2 /e v4 ( x ) = p ( a e e3 1
1 4 4
a ( x c1 ) 4
v2 ( x ) = p ( a e e3 1
4
a | x + c1 |
2 2 3/2
+ a4 ei/4 e 3 i a(xc1 ) /e )(1 + O( x c1 )),
4
2 3/2
+ a2+ ei/4 e 3 i a| x+c1 | /e )(1 + O( x + c1 )),
4
x c1 ,
x c1 , 1
c3 sin( 23 |z|3/2 +

v 34 ( x ) = p 4)
1 4
2 |z|
c2 sin( 23 | |3/2

v 23 ( x ) = p + 4 )
4
2 | | + c3+ cos( 23 |z|3/2 + 4 ) (1 + O(1/|z|3/2 )),

z .
+ c2+ cos( 23 | |3/2 + 4 ) (1 + O(1/| |3/2 )), .

Matching these expansions, we obtain
 
( ae)1/6 1 i
 +
Requiring that these two expansions match, we obtain the con-
 
a4 c3
= C7 + , C7 = .
nection formula a4 c3 2 2 1 i
4
 +  + 4
2
 
c2 a2 1 1
= C4 , C4 = . Note also that a4 in Eq. (31b) relates to a4 in Eq. (31a) by
c2 a2 ( ae)1/6 i i  +  +
a4 a4
= C .
Next we match v3 ( x ) with v23 ( x ). Note that a4 3
a4
1 2 3/2 Then we match v4 ( x ) with v45 ( x ). Note that
( a3 e 3 a(x+c1 ) /e
4
v3 ( x ) = p
4
a ( x + c1 ) 1 i/4 2 ia| xc |3/2 /e
4
2 3/2 v4 ( x ) = p ( a 4 e e3 2

+ a3+ e 3 a(x+c1 ) /e )(1 + O( x + c1 )),

4 4
a | x c2 |
2 3/2
x c1 , + a4+ ei/4 e 3 i a| xc2 | /e )(1 + O( x c2 )),
4

1 2 3/2 2 3/2 x c2 ,
v 23 ( x ) = p ( 12 c2 e 3 + c2+ e 3 )
4 2
1
c4 sin( 23 ||3/2 +

3/2 v 45 ( x ) = p 4)
(1 + O(1/ )), . 4
2 ||
+ c4+ cos( 23 ||3/2 + 4 ) (1 + O(1/||3/2 )), .

Matching these expansions, we obtain
   +
( ae)1/6 1 0
  Requiring that these two expansions match, we obtain the con-
a3 c2
= C5 , C5 = 4 1 . nection formula
a3+ c2 2 0 2  +  + 4
2 i i
 
c4 a4
= C , C = .
Note that a3 in Eq. (30b) relates to a3 in Eq. (30a) by c4 8
a4 8
( ae)1/6 1 1
   
a3 a Finally we math v5 ( x ) with v45 ( x ). Note that
= C6 3+ , C6 = e3 S2 ()/e
a3+ a3 1 2 3/2
( a5 e 3 a(xc2 ) /e
4
v5 ( x ) = p
4
with a ( x c2 )
2 3/2
+ a5+ e 3 a(xc2 ) /e )(1 + O( x c2 )),
4
Rc3 q
S2 ( ) = | Q( x )|dx. (33)
c2 x c2 ,
1 2 3/2 2 3/2
A plot of S2 () is shown in Fig. 7. v 45 ( x ) = p
4
( 12 c4 e 3 + c4+ e 3 )
2

To match v3 ( x ) with v34 ( x ), note that
1 2 a| xc |3/2 /e (1 + O(1/3/2 )), .
+ 4
v3 ( x ) = p ( a 3 e 3 1
4
a | x c1 | Matching these expansions, we obtain
2 3/2
 
a5
 +
c4
+ a3 e 3 a| xc1 | /e )(1 + O(| x c1 |)),
4
= C .
a5+ 5
c4
x c1 ,

2 2
1.5 1.5
S1
S2
1 1
0.5 0.5
0 0
0 0.5 1 0
FIG. 7. Left: S1 () as a function of (red curve). The gray (black)
curve shows the linear (quadratic) approximation of S1 () about
= 1. Right: S2 () as a function of (red curve). The gray
curve shows the linear approximation of S2 () about = 1.
Trace of the monodromy matrix for 0 < < 1. We now
finally ready to use the results of the previous section to com-
pute the monodromy matrix for 0 < < 1. In the sub-region
x (, c1 ), we choose the fundamental matrix solution
Y ( x ) as in Eq. (24), with v+ ( x ) and v ( x ) given by Eq. (27).
The value of this fundamental matrix solution at x = is
thus
Y ( ) =
 
S ()/2e
1
A ( )e 2 S A+
e | Q( )| A ( )e 2 ()/2e 1e | Q( )| A+ ( )eS2 ()/2e
with A given by Eq. (28). Based on the above discussion,
the value of the continuation of the above fundamental matrix
solution at x = is given by
Y ( ) =
 
S ()/2e
A ( )e 2 S ()/2e A + ( )e
1
e | Q( )| A ( )e 2 1e | Q( )| A+ ( )eS2 ()/2e
where C is the connnection matrix given by
C = C5 C8 C3 C7 C1 C6 C5 C4 C3 C2 C1 .
Now recall that, as before, we can obtain the monodromy ma-
trix via Eq. (25). Tedious but straightforward algebra then
shows that the trace of the monodromy matrix is given by
Eq. (7), with S1 () and S2 () given by Eq. (32) and Eq. (33),
as before.
A.3 Calculation of relative bandwidth
Approximated expressions for S1 () and S2 (). The
functions S1 () and S2 () characterise the oscillating and
the envelope part of the analytical expression of the trace of
the monodromy matrix in Eq. (25). These are crucial quanti-
ties since S1 () determine the location of the bands through
Eq. (9) and in turn the amplitudes and velocities of the rel-
ative soliton-like excitations, whereas the solitonic character
of the band is fixed by the condition (11) which depends on
S2 (). In order to obtain explicit estimates of the number of
10
solitons along with their features (amplitudes and velocities)
as a function of e, it is worth to introduce appropriate Tay-
lor expansions. Specifically, the Taylor expansions of S1 ()
around = 0 and of S2 () around = 1 are needed. Using
known properties of elliptic integrals [63], one can verify that
these expansions are given by the following expressions
S1 () = /2 + 22 /16 + O(3 ) , (35a)
0.5 1
S2 () = ( 1)/2 + O( 1)2 , (35b)
respectively.
Relative bandwidth. Here we give some details on the
characterization of the bands. For n 1, the width of the
n-th spectral band, which is approximately centered at zn , is
given by wn = 2n 2n1 , while for n = 0, the bandwidth
is given by w0 = 0 . The width of the n-th spectral gap is
given by gn = 2n+1 2n . We first consider the Taylor ex-
pansion of trM around zn (the location of the n-th maximum),
evaluated at = 2n1 and = 2n , and we obtain the n-th
bandwidth as
2e
wn = exp(S2 (zn )/e) + O(ee2S2 (zn )/e ) .
|S10 (zn )|
Next we use the difference zn+1 zn to approximate wn +
gn . To do so, we first show the difference between these two
( )eS2 ()/2e
, quantities is given by
(wn + gn ) (zn+1 zn ) = O(eeS2 (zn+1 )/e ).
We then expand S1 () around zn , evaluate at = zn+1 and
obtain
zn+1 zn = e/S10 (zn ) + O(e2 ).
S2 ()/2e
C, Combining all the above results, we finally obtain the asymp-
totic expression for the n-th relative bandwidth in Eq. (10).
(34)
[1] E. Fermi, J. Pasta and S. Ulam, Studies of nonlinear problems,
in Collected Papers of Enrico Fermi, E. Segre, Ed. (University
of Chicago, Chicago, 1965), Vol. 2.
[2] J.L. Tuck and M.T. Menzel, Adv. Math. 9, 399 (1972).
[3] J. Ford, Phys. Rep. 213, 271-310 (1992).
[4] G. P. Berman and F. M. Izrailev, Chaos 15, 015104 (2005).
[5] T. Dauxois, M. Peyrard and S. Ruffo, Eur. J. Phys. 26, S3
(2005).
[6] G. Gallavotti (ed.) The Fermi-Pasta-Ulam problem: a status re-
port, Lect. Note Phys. 728 (Springer, Berlin-Heidelberg, 2008).
[7] M. A. Porter, N. J. Zabusky, B. Hu and D. K. Campbell, Am.
Sci. 97, 214 (2009).
[8] G. Benettin, H. Christodoulidi, A. Ponno, J. Stat. Phys. 152,
195 (2013).
[9] M. Onorato, L. Vozella, D. Proment, and Y. V. Lvov, PNAS
112, 4208 (2015).
[10] C. Bao, J. A. Jaramillo-Villegas, Y. Xuan, D. E. Leaird, M. Qi,
and A. M. Weiner, Phys. Rev. Lett. 117, 163901 (2016).
[11] M. Guasoni, J. Garnier, B. Rumpf, D. Sugny, J. Fatome, F. Am-
rani, G. Millot, and A. Picozzi, Phys. Rev. X 7, 011025 (2017).

[12] N. Zabusky and M. Kruskal, Phys. Rev. Lett. 15, 240243
(1965).
[13] J. Korteweg and G. De Vries, Phil. Mag. 39, 422 (1895).
[14] P. Lorenzoni and S. Paleari, Phisica D 221, 110 (2006).
[15] M. Toda, J. Phys. Soc. Jpn. 26, 235 (1969).
[16] M. Toda, Theory of nonlinear lattices (Springer, New York,
1981).
[17] R. Hirota and K. Suzuki, J. Phys. Soc. Jpn. 28, 1366 (1970).
[18] H. Ikezi, Phys. Fluids 16, 1668 (1973).
[19] S. Trillo, G. Deng, G. Biondini, M. Klein, G. F. Clauss, A.
Chabchoub and M. Onorato, Phys. Rev. Lett. 117, 144102
(2016).
[20] Y. C. Mo, R. A. Kishek, D. Feldman, I. Haber, B. Beaudoin, P.
G. OShea, and J. C. T. Thangaraj, Phys. Rev. Lett. 110, 084802
(2013).
[21] W. E. Ferguson, H. Flaschka, and D. W. McLaughlin, J. Comp.
Phys. 45, 157 (1982).
[22] N.M. Ercolani, I. R. Gabitov, C. D. Levermore, and D. Serre,
eds., Singular Limits of Dispersive Waves, NATO ASI Series,
Ser. B 320 (Plenum Press, New York, 1994).
[23] G. A. El and M. A. Hoefer, Physica D 333, 11 (2016).
[24] Z. Dutton, M. Budde, C. Slowe, L. V. Hau, Science 293, 663
(2001).
[25] M. A. Hoefer, M. J. Ablowitz, I. Coddington, E. A. Cornell, P.
Engels, and V. Schweikhard, Phys. Rev. A 74, 023623 (2006).
[26] W. Wan, S. Jia, and J. W. Fleischer, Nature Phys. 3, 46 (2007).
[27] N. Ghofraniha, C. Conti, G. Ruocco, and S. Trillo, Phys. Rev.
Lett. 99, 043903 (2007).
[28] C. Conti, A. Fratalocchi, M. Peccianti, G. Ruocco, and S. Trillo,
Phys. Rev. Lett. 102, 083902 (2009).
[29] S. Trillo and A. Valiani, Opt. Lett. 35, 3967 (2010).
[30] J. Fatome, C. Finot, G. Millot, A. Armaroli and S. Trillo, Phys.
Rev. X 4, 021022 (2014).
[31] A. Moro and S. Trillo, Phys. Rev. E 89, 023202 (2014).
[32] G. Xu, M. Conforti, A. Kudlinski, A. Mussot, and S. Trillo,
Phys. Rev. Lett. 118, 254101 (2017).
[33] P. A. P. Janantha, P. Sprenger, M. A. Hoefer, and M. Wu, Phys.
Rev. Lett. 119, 024101 (2017).
[34] P. D. Miller and Z. Xu, Commun. Pure Appl. Math. 64, 205
(2010).
[35] J. Garnier, G. Xu, S. Trillo, and A. Picozzi, Phys. Rev. Lett.
111, 113902 (2013).
[36] B. L. Holian, H. Flaschka and D. W. McLaughlin, Phys. Rev. A
24, 2595 (1981).
[37] S. Venakides, P. Deift, and R. Oba, Commun. Pure Applied
Math. 44, 1171 (1991).
11
[38] R. F. Bibkaev, J. Math. Sciences 77, 3033 (1995).
[39] B. M. Lake, H. C. Yuen, H. Rungaldier and W. E. Ferguson, J.
Fluid Mech. 83, 49 (1977).
[40] H. C. Yuen and W. E. Ferguson, Phys. Fluids 21, 1275 (1978).
[41] N. N. Akhmediev, V.M. Eleonoskii, and N.E. Kulagin, Sov.
Phys. JETP 62, 894 (1985).
[42] S. Trillo and S. Wabnitz, Opt. Lett. 16, 986 (1991).
[43] G. Van Simaeys, Ph. Emplit, and M. Haelterman, Phys. Rev.
Lett. 87, 033902 (2001).
[44] O. Kimmoun, H.C. Hsu, H. Branger, M.S. Li, Y.Y. Chen, C.
Kharif, M. Onorato, E. J. R. Kelleher, B. Kibler, N. Akhmediev,
and A. Chabchoub, Sci. Rep. 6, 28516 (2016).
[45] B. Dubrovin, T. Grava, and C. Klein, J. Nonlinear Sci. 19, 57
(2009).
[46] P. D. Miller and S. Kamvissis, Phys. Lett. A 247, 75 (1998).
[47] S. Kamvissis, K. T. R. McLaughlin, and P. D. Miller, Semiclas-
sical soliton ensembles for the focusing nonlinear Schrodinger
equation, (Princeton University Press, Princeton, 2003).
[48] M. J. Ablowitz and H. Segur, Solitons and the inverse scattering
transform (SIAM, Philadelphia, 1981)
[49] E. Infeld and G. Rowlands, Nonlinear waves, solitons and
chaos (Cambridge University Press, Cambridge, 2000)
[50] G. P. Agrawal, Nonlinear fiber optics (Academic Press, New
York, 2007)
[51] J. R. Thompson and R. Roy, Phys. Rev. A 43, 4987 (1991).
[52] S. Trillo, S. Wabnitz, and T. A. B. Kennedy, Phys. Rev. A 50,
1732 (1994).
[53] A.R. Its and V.P. Kotlyarov, Dokl. Akad. Nauk Ukr. RSR A 11,
965 (1976).
[54] Y. Ma and M. Ablowitz, Stud. Appl. Math. 65, 113 (1981).
[55] E.D. Belokolos, A.I. Bobenko, V.Z. EnolSkii and A.R. Its,
Algebro-geometric approach to nonlinear integrable equations
(Springer, New York, 1994).
[56] C. S. West and T. A. B. Kennedy, Phys. Rev. A 47, 1252 (1993).
[57] A. Messiah, Quantum mechanics, Vol. I (Elsevier, Amsterdam,
1961)
[58] G. Deng, G. Biondini and S. Trillo, Physica D 333, 137 (2016).
[59] A.R. Osborne and L. Bergamasco, Physica D 18, 26 (1986).
[60] A. Fratalocchi, C. Conti, G. Ruocco, and S. Trillo, Phys. Rev.
Lett. 101, 044101 (2008).
[61] P. V. Mamyshev, P. G. Wigley, J. Wilson, C. Bosshard, and G.I.
Stegeman, Appl. Phys. Lett. 64, 3374 (1994)
[62] P. V. Mamyshev, C. Bosshard, and G.I. Stegeman, J. Opt. Soc.
Am. B. 11, 1254 (1994).
[63] F. W. Olver, D. W. Lozier, R. F. Boisvert and C. W. Clark, NIST
Handbook of Mathematical Functions, (Cambridge University
Press, New York, 2010)